OpenAI Publishes 722 AI-Generated Math Papers Without Disclosing Methods

OpenAI released 722 mathematics manuscripts on October 6, grouped into 372 families of related results across 17 mathematical fields, generated by an unnamed internal reasoning model.

The release included formal Lean proof verifications for many manuscripts and reasoning summaries for ten result families, but withheld critical details:

• The prompts used to generate results were not disclosed

• The model's name and version remained unidentified

• Compute resources used were not specified

Mathematicians on MathOverflow and elsewhere raised concerns about:

• Whether results properly credit or build on specific pre-existing human research

• The difficulty of auditing outputs when underlying prompts are undisclosed

• A related credit dispute with NYU mathematician Tristan Buckmaster concerning an earlier Navier-Stokes result from September 8

The release was published to a public GitHub repository under an Apache 2.0 license, making the manuscripts freely available — but the lack of methodological transparency has made it difficult for the research community to verify where AI reasoning started and where prior human work left off.

Why It Matters: The dispute highlights a genuine open problem in AI-assisted science: without disclosed prompts or clear citation trails, the research community cannot reliably audit AI-generated mathematical claims.